pulseml 0.2.1__tar.gz → 0.2.2__tar.gz

{pulseml-0.2.1 → pulseml-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pulseml
-Version: 0.2.1
+Version: 0.2.2
 Summary: A live ML training debugger - GUI or CLI, any backend (NumPy, PyTorch, TensorFlow, CuPy, JAX).
 Author-email: Yash Patel <codeyash09@gmail.com>
 License: Proprietary

{pulseml-0.2.1 → pulseml-0.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "pulseml"
-version = "0.2.1"
+version = "0.2.2"
 description = "A live ML training debugger - GUI or CLI, any backend (NumPy, PyTorch, TensorFlow, CuPy, JAX)."
 readme = "README.md"
 requires-python = ">=3.9"

{pulseml-0.2.1 → pulseml-0.2.2}/src/pulse/pulse.py

@@ -3541,6 +3541,24 @@ def _install_cli_excepthook(cli):
     previous_hook = sys.excepthook
 
     def _hook(exc_type, exc_value, exc_tb):
+        # Disarm the global line/call tracer immediately. Once the script
+        # has crashed, nothing below should still be traced -- and if we
+        # leave `sys.settrace` pointed at `cli_tracer`, it stays the
+        # process-wide trace function into interpreter shutdown, where it
+        # gets invoked on unrelated __del__ calls (e.g. litellm's
+        # AsyncHTTPHandler) after module globals have already been cleared
+        # to None. That produces a second, uncatchable "Exception ignored
+        # in ..." failure from *inside* the tracer, separate from (and
+        # after) whatever we report here.
+        sys.settrace(None)
+        try:
+            frame = exc_tb.tb_frame if exc_tb else None
+            while frame is not None:
+                frame.f_trace = None
+                frame = frame.f_back
+        except Exception:
+            pass
+
         previous_hook(exc_type, exc_value, exc_tb)
         if issubclass(exc_type, KeyboardInterrupt):
             return
@@ -3634,14 +3652,29 @@ def _start_cli_tracker(
     cli.extra_files = dict(extra_files or {})
     cli.pending_startup_error = startup_error
     cli.print_banner()
-    _install_cli_excepthook(cli)
+    # Agent/provider setup used to be deferred until cli_tracer saw its
+    # first 'line' event with a trackable/resolved local in scope -- fine
+    # for a training loop, but it meant any crash before that point (e.g.
+    # failing during data loading, before the loop even starts) left
+    # cli.agent_provider unset, so the crash hook below had no agent to
+    # hand the traceback to and could only print "no AI agent is
+    # configured". discover_variables() already falls back to the
+    # statically-discovered `cli.discovered` names (populated from AST
+    # parsing at auto_track()-call time, before any user code runs) when
+    # there are no live locals yet, so setup doesn't actually need to wait
+    # for a live frame -- run it eagerly, before the excepthook can matter.
+    cli.interactive_setup()
 
     if startup_error:
         print("[Pulse] The function passed to auto_track() raised an exception during its dry run:")
         print(startup_error)
-        print("[Pulse] Set up an AI agent below and Pulse will offer to diagnose/fix it before continuing.\n")
+        if cli.agent_provider:
+            print("[Pulse] Pulse will offer to diagnose/fix it (via the agent set up above) before continuing.\n")
+        else:
+            print("[Pulse] No agent configured (see summary above) -- run /agent to set one up before continuing.\n")
+
+    _install_cli_excepthook(cli)
 
-    setup_done = {"value": False}
     last_logged = {"t": 0.0}
 
     # ------------------------------------------------------------------
@@ -3672,7 +3705,7 @@ def _start_cli_tracker(
     # next window. Between windows, the training loop runs at native,
     # untraced speed.
     _CAPTURE_SPAN = min(0.05, throttle_interval / 4 if throttle_interval > 0 else 0.05)
-    window = {"open": True, "closes_at": 0.0}  # start open so setup can run immediately
+    window = {"open": True, "closes_at": 0.0}  # start open so the first snapshot can run immediately
 
     def _close_window():
         window["open"] = False
@@ -3694,6 +3727,11 @@ def _start_cli_tracker(
             if window["open"]:
                 _close_window()
 
+    # Setup now runs eagerly, before tracing even starts (see above), so
+    # the ticker can start right away too -- no need to wait for the
+    # tracer to see a live frame first.
+    threading.Thread(target=_ticker, daemon=True).start()
+
     def cli_tracer(frame, event, arg):
         filename = frame.f_code.co_filename
         if _is_library_frame(filename):
@@ -3740,21 +3778,6 @@ def _start_cli_tracker(
                 ):
                     cli.var_states[name] = "track"
 
-        if not setup_done["value"]:
-            has_resolved_shape = any(shape is not None for shape in cli.discovered.values())
-            has_trackable_local = any(
-                not n.startswith("__") and is_trackable(v) for n, v in local_vars.items()
-            )
-            if has_resolved_shape or has_trackable_local:
-                cli.watch_locals = local_vars
-                cli.interactive_setup()
-                setup_done["value"] = True
-                # Setup is interactive/blocking; once it's done, start the
-                # background ticker that opens/closes future windows. (Kept
-                # off until now so setup itself isn't racing a window close.)
-                threading.Thread(target=_ticker, daemon=True).start()
-            return cli_tracer
-
         cli.watch_locals = local_vars
 
         now = time.time()

{pulseml-0.2.1 → pulseml-0.2.2}/src/pulseml.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pulseml
-Version: 0.2.1
+Version: 0.2.2
 Summary: A live ML training debugger - GUI or CLI, any backend (NumPy, PyTorch, TensorFlow, CuPy, JAX).
 Author-email: Yash Patel <codeyash09@gmail.com>
 License: Proprietary